Boosted Deep Convolutional Neural Networks Training

The patent divides the complex neural network into multiple stages of training with different architectures. Initially, a simpler network is trained to establish basic classification capabilities, then progressively more complex networks are trained and combined. This segmentation allows the system to achieve high accuracy without requiring all complexity to be present simultaneously, resolving the contradiction between accuracy and complexity.

The patent implements preliminary training actions using simpler networks before deploying complex architectures. By pre-training with reduced complexity and progressively adding complexity in subsequent training stages, the system achieves high classification accuracy while managing device complexity through staged introduction of complex elements rather than immediate full complexity.

The training process is segmented into multiple stages, each using different dataset sizes and network complexities. Instead of requiring all tens of millions of training samples from the outset, the system progresses through stages where earlier stages use smaller datasets with simpler networks, then combines these with additional training data in later stages. This segmentation dramatically reduces the time required to achieve high accuracy compared to traditional approaches that require full dataset training from the beginning.

Preliminary training actions are performed with smaller datasets and simpler architectures to establish baseline classification capabilities quickly. These preliminary results are then combined with additional training data in subsequent stages, allowing the system to achieve high classification accuracy without requiring the full extent of time-consuming training with all training samples simultaneously.